Seeing Inside the Black Box: Enhancing Trust in Brain Tumour Detection

by

Author: Denis Avetisyan

New research demonstrates how combining multiple explainable AI techniques can provide a more complete and trustworthy understanding of deep learning models used in medical imaging.

Integrating SHAP, GRAD-CAM, and LRP yields a robust framework for interpreting convolutional neural network decisions in brain tumour detection.

Despite advances in deep learning for medical image analysis, understanding why these models make specific diagnoses remains a significant challenge. This is addressed in ‘Explainable AI: A Combined XAI Framework for Explaining Brain Tumour Detection Models’, which proposes an integrated approach to enhance the interpretability of convolutional neural networks for brain tumour detection. By combining Gradient-weighted Class Activation Mapping (GRAD-CAM), Layer-wise Relevance Propagation (LRP), and SHapley Additive exPlanations (SHAP), this research demonstrates superior explanatory power compared to individual methods, offering insights from broad spatial regions to pixel-level details. Could this multi-technique framework foster greater trust and facilitate clinical adoption of AI-driven diagnostic tools in neuro-oncology?

Decoding the Subtle Signals: The Challenge of Brain Tumor Detection

The detection of brain tumors presents a significant clinical challenge, not simply because of their presence, but due to their remarkable heterogeneity. Tumors exhibit substantial variations in both appearance and location within the complex anatomy of the brain. This inconsistency stems from diverse cellular compositions, differing growth patterns, and the unique physiological environment of each patient. Consequently, a tumor that appears diffuse and ill-defined in one individual may present as a relatively compact and sharply-bordered mass in another. Furthermore, tumors situated deep within the brain, or those bordering critical structures, can be particularly difficult to visualize and differentiate from surrounding healthy tissue. This inherent complexity necessitates advanced diagnostic techniques capable of overcoming these limitations and ensuring accurate identification, even in the face of subtle or atypical presentations.

The accurate and timely identification of brain tumors is frequently hampered by the challenge of partial tumor visibility during medical imaging. Conventional diagnostic techniques can struggle when tumors are located near critical brain structures, obscured by surrounding tissue, or exhibit diffuse margins, leading to underestimation of tumor size or even complete misdiagnosis. This limitation has significant clinical consequences, as delayed or inaccurate assessments can postpone necessary treatment interventions, potentially allowing the tumor to progress and negatively impacting patient outcomes. The difficulty arises because standard image analysis often relies on clearly defined boundaries, which are absent in cases of partial visibility, necessitating more sophisticated approaches to reliably detect and delineate these subtle anomalies.

Despite the widespread use of FLAIR (Fluid-Attenuated Inversion Recovery) magnetic resonance imaging for its high sensitivity in detecting edema and lesions indicative of brain tumors, inherent limitations necessitate the application of robust analytical techniques. While FLAIR excels at suppressing cerebrospinal fluid signal, enhancing the visibility of pathological changes, factors such as signal heterogeneity within tumors, varying degrees of peritumoral edema, and susceptibility artifacts can obscure clear delineation of tumor boundaries. Consequently, relying solely on visual inspection of FLAIR images often proves insufficient for precise tumor characterization and volumetry. Advanced computational methods, including machine learning algorithms and quantitative image analysis, are therefore crucial for extracting subtle features, normalizing signal intensities, and ultimately improving the accuracy of tumor detection and monitoring, even within the sensitive context of FLAIR imaging.

Engineering Precision: A CNN for Tumor Identification

The custom Convolutional Neural Network (CNN) architecture employed for brain tumor detection and classification consists of five convolutional blocks, each incorporating 2D convolutional layers with ReLU activation functions, followed by max-pooling layers for dimensionality reduction. These blocks extract hierarchical features from the input MRI scans. A final fully connected layer, utilizing a softmax activation function, outputs the probability distribution across the defined tumor classes: meningioma, glioma, pituitary tumor, and normal tissue. The architecture was designed with a specific focus on capturing subtle image features indicative of tumor presence and type, differing from general-purpose CNNs through optimized filter sizes and layer configurations determined through experimentation with the BraTS 2021 dataset.

Robust image preprocessing was critical to the performance of the convolutional neural network (CNN) when analyzing MRI scans from the BraTS 2021 dataset. This preprocessing pipeline included intensity normalization to a standard range, bias field correction to account for scanner inhomogeneities, and resampling to a consistent isotropic voxel size of 1x1x1 mm3. Skull stripping was performed to remove non-brain tissue, and data augmentation techniques-including random rotations, flips, and scaling-were employed to increase the dataset’s variability and prevent overfitting. These steps standardized the input data, reducing noise and improving the CNN’s ability to accurately identify and classify tumor regions.

Model training prioritized both high tumor detection accuracy and the minimization of false positive results, resulting in an overall accuracy of 91.24% when evaluated against the BraTS 2021 Dataset. This performance metric was achieved through iterative optimization of training parameters and network architecture. Comparative analysis demonstrates a statistically significant improvement over previously published results on the BraTS 2021 Dataset, indicating enhanced reliability in identifying and classifying brain tumors. The model’s ability to reduce false positives is particularly crucial for minimizing unnecessary patient interventions and improving diagnostic confidence.

Unveiling the Logic: XAI for Trustworthy Diagnosis

To mitigate the inherent lack of transparency in Convolutional Neural Networks (CNNs), we incorporated Explainable AI (XAI) methods – specifically Gradient-weighted Class Activation Mapping (GRAD-CAM), Layer-wise Relevance Propagation (LRP), and SHapley Additive exPlanations (SHAP). GRAD-CAM generates heatmaps highlighting image regions influencing the CNN’s output, while LRP traces the prediction back to the input features, quantifying their contribution. SHAP, utilizing game theory, assigns each feature an importance value for a particular prediction. Integrating these techniques allows for the visualization and quantification of the model’s decision-making process, facilitating understanding and trust in the CNN’s predictions.

The integration of Explainable AI (XAI) techniques facilitates the generation of heatmaps overlaid on the input MRI scans. These visualizations highlight the specific image regions that most strongly contributed to the model’s diagnostic prediction. This capability is critical for clinical validation, allowing radiologists and other healthcare professionals to assess whether the model is focusing on clinically relevant anatomical features or potential artifacts. By providing this granular level of insight into the model’s decision-making process, clinicians can gain increased confidence in the AI’s output and more effectively integrate it into their diagnostic workflow, ultimately supporting improved patient care.

Performance metrics demonstrate a significant improvement with the implemented model. Precision reached 96.08% and recall achieved 86.22%, representing gains of 4.54% and 13.12% over the baseline model, respectively. The F1-score, a harmonic mean of precision and recall, increased to 90.88%, a 9.06% improvement, indicating a more balanced and reliable predictive capability.

Beyond Prediction: Towards a Collaborative Diagnostic Future

The development of an Explainable AI (XAI)-enhanced Convolutional Neural Network (CNN) signifies a promising advancement toward more effective diagnostic tools for medical professionals. This system isn’t intended to replace clinicians – that would be a crude simplification – but rather to function as a powerful decision support mechanism, augmenting their expertise with detailed visual explanations. By highlighting the specific features within medical images that contribute to a diagnosis, the XAI-enhanced CNN allows for greater scrutiny and validation of AI-driven insights. This capability is particularly crucial in complex cases, where subtle anomalies might be overlooked, and supports more informed treatment planning by providing a transparent rationale behind the model’s conclusions. Ultimately, this technology aims to foster a collaborative environment where clinicians and AI work in synergy to improve patient outcomes.

The enhanced diagnostic model moves beyond simply providing a diagnosis; it actively supports clinical judgment through visual explanations. These visualizations highlight the specific features within medical images that contributed to the model’s conclusions, enabling clinicians to validate the AI’s reasoning and build confidence in its assessments. This capability is particularly crucial for identifying subtle anomalies – those often missed by the human eye – and refining treatment strategies. By illuminating the ‘why’ behind the diagnosis, the model facilitates a more collaborative approach, empowering medical professionals to leverage AI as a powerful extension of their expertise, rather than a black box, and ultimately leading to improved patient outcomes.

The newly developed diagnostic model exhibits substantially improved performance, as evidenced by a test loss of 0.2355 – a noteworthy reduction from the original model’s 0.3482. This decrease in loss not only signifies enhanced accuracy but also demonstrates greater stability and the ability to generalize effectively to unseen data. Such advancements are crucial for building trust in artificial intelligence within healthcare settings, as clinicians require reliable and understandable tools. The resulting system is poised to function as a collaborative partner, empowering medical professionals to validate findings, pinpoint subtle indicators, and refine treatment strategies with increased confidence and precision.

The pursuit of reliable brain tumour detection, as detailed in this research, isn’t merely about achieving high accuracy; it’s about dissecting how that accuracy is attained. This aligns perfectly with Robert Tarjan’s assertion: “Program verification is essentially the art of constructing infallible arguments.” The combined XAI framework-integrating GRAD-CAM, LRP, and SHAP-functions as such an argument, providing a multi-faceted ‘proof’ of the model’s reasoning. By layering these interpretability techniques, the system moves beyond simply identifying a tumour to revealing which features drive that identification, and thus, verifying the validity of its conclusions. It’s a controlled demolition of the ‘black box’, revealing the underlying logic-a testament to understanding through rigorous examination.

Beyond the Black Box

The integration of XAI techniques-as demonstrated by this work-isn’t simply about seeing what a convolutional neural network attends to; it’s about acknowledging that ‘attention’ is itself a construct. The model doesn’t ‘reason’ about a brain tumour; it correlates patterns. Combining GRAD-CAM, LRP, and SHAP offers a more robust illusion of understanding, but the fundamental opacity remains. Future work should, therefore, move beyond merely illuminating the decision process and begin to actively stress-test it. What subtle perturbations in input data cause the most dramatic shifts in explanation, and how reliably do these shifts correlate with actual clinical significance?

True security, in this context, isn’t found in increasingly complex explanation methods, but in brutally honest assessments of their limitations. The field risks becoming enamored with elegant visualizations that offer little genuine insight into model vulnerabilities. A fruitful avenue lies in quantifying the disagreement between different XAI methods – a high degree of inconsistency isn’t a bug, but a flag indicating areas where the model’s logic is least stable, and therefore, most deserving of scrutiny.

Ultimately, the goal shouldn’t be to explain the model to the clinician, but to equip the clinician with tools to interrogate it. The physician must remain the final arbiter, not a passive recipient of algorithmic pronouncements, however neatly visualized. The path forward requires not just more sophisticated XAI, but a shift in paradigm: from explanation to controlled deconstruction.

Original article: https://arxiv.org/pdf/2602.05240.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-02-07 14:37